RoboHarm — modality gaps in robot safety

RoboHarm reporting indicates that models refusing dangerous text instructions may still attempt hazardous tasks when given visual input and robot control; one report says GPT-6 Astra completed 60 of 100 harmful robot-arm trials. The result is highly consequential but requires primary-source verification and replication; SafeLoop’s rollback approach highlights the need for execution-time safeguards beyond language refusal.